More like, closed source models are not actually better in any meaningful sense and people and companies can become extremely rational when it comes to spending money.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
I find the fact that DS still calls their SOTA models V3 hilarious.
Keep in mind this is written as the intro to their SOTA model, which they still calls V3.
You usually outline the problem that your paper solves in the intro.
I also find it funny that Chinese open source models are so good people just automatically assume what's released must be their latest and there were no delay from completion to release. Unlike closed source models they're not making money on it so there's nothing to be lost by holding back and releasing it at the optimal time for street cred, especially if you're also a hedge fund....
Closed source models from top labs are still meaningfully better. If you've tried them side by side, you'd know this. I've actually sat and rated many generations from e.g. Nano Banana 2 vs. Qwen Edit, and yes Nano Banana 2 remains superior (mainly due to its multi-modal understanding capabilities and presumably much larger knowledge base, which is reflected in its model size). But the issue is that Nano Banana 2 is 1) expensive, 2) highly censored, and 3) not customizable in any way, since it's a proprietary API service. This makes it much less attractive to build around compared to Chinese models.
The race is still on and even though I'm optimistic, I can see why Deep Seek would claim that closed models are accelerating their rate of improvement. Infinite money and scale does have some benefits in enabling thousands of parallel experiments, top tier AI talent (many poached from China), and data gathering / labeling / collection at a rate that start-ups like Deep Seek, Z.AI, and Moon Shot would find hard to match. It's why Western labs still have the lead in frontier performance, but the industry is learning that often times, you don't need frontier performance and it isn't necessarily worth the cost, either.
I do compare them side by side between work and personal and no they're not meaningfully better for what I do.
And that's when I don't have to pay for the former.
Or perhaps because I don't pay for the former I don't have a reason to convince myself it's better by telling myself it has presumably larger knowledge base.
Way too early and premature to be saying that. We don’t even know what will happen tomorrow much less what will happen years from now.
my contact Taylor put together a thread of trying out the Doubao phone. It's really pretty cool.